A Parameter Communication Optimization Strategy for Distributed Machine Learning in Sensors.

Area of Science:

• Distributed machine learning
• Sensor networks
• Optimization strategies

Background:

• Distributed machine learning leverages sensor networks but faces challenges from varying sensor capabilities and network delays.
• These factors significantly impact machine learning model accuracy and convergence rates.
• Existing methods struggle to balance training and communication overhead effectively.

Purpose of the Study:

• To propose a parameter communication optimization strategy for distributed machine learning.
• To enhance the accuracy and convergence rate of machine learning models in sensor networks.
• To balance training and communication overhead in distributed systems.

Main Methods:

• Developed Dynamic Finite Fault Tolerance (DFFT) to extend fault tolerance for iterative algorithms.
• Implemented the Dynamic Synchronous Parallel Strategy (DSP) based on DFFT.
• Utilized a performance monitoring model for dynamic adjustment of parameter synchronization between worker nodes and Parameter Servers (PS).

Main Results:

• The DSP strategy effectively utilizes sensor computing power.
• Ensured the accuracy of the machine learning model.
• Prevented model training disruptions from unrelated tasks.

Conclusions:

• The proposed Dynamic Synchronous Parallel Strategy (DSP) offers an effective solution for optimizing distributed machine learning in sensor networks.
• DSP enhances model performance by dynamically managing parameter synchronization.
• This approach improves the robustness and efficiency of distributed machine learning systems.